LIP. Laboratoire de l Informatique du Parallélisme. Ecole Normale Supérieure de Lyon Institut IMAG Unité de recherche associée au CNRS n 1398

Transcription

1 LIP Laboratoire de l Informatique du Parallélisme Ecole Normale Supérieure de Lyon Institut IMAG Unité de recherche associée au CNRS n 1398 On the redundancy of real number representation systems Christophe Mazenc May 1993 Research Report N o Ecole Normale Supérieure de Lyon 46, Allée d Italie, Lyon Cedex 07, France, Téléphone : ; Télécopieur : ; Adresses électroniques : lip@frensl61.bitnet; lip@lip.ens lyon.fr (uucp).

2 On the redundancy of real number representation systems Christophe Mazenc May 1993 Abstract In this paper, a set of denitions describing general real number representation systems is presented. Our purpose is to nd a suciently wide model denition including classical systems (signed-digit notation, linear numeration systems, p-adic numbers, symmetric level index arithmetic...) but specic enough to make it possible to build some practical results. We focuse on the redundancy property and the relationships between redundancy, on-line and digit-parallel calculus. Keywords: Number representation systems, redundancy, on-line arithmetic Resume Dans ce papier, nous presentons un ensemble de denitions decrivant de maniere generale tout systeme de representation des nombres. Notre but est de denir un modele suf- samment general pour inclure les systemes classiques et susamment specique pour rendre possible la synthese de proprietes generales. Nous nous interessons plus particulierement a la propriete de redondance et explicitons ses liens avec les modes de calcul en ligne et parallele au niveau du chire. Mots-cles: Systemes de representation des nombres, redondance, arithmetique en ligne

3 On the redundancy of real number representation systems y Christophe Mazenc LIP Ecole Normale Superieure de Lyon 46 Allee d'italie Lyon Cedex 07, France Abstract A lot of people whose research eld has a non empty intersection with computer arithmetic take for granted many basic assumptions such as : \if we want to perform serial addition most signicant digits rst, we have to use a redundant number system" or \if the system is not redundant, one can not build a fully digit parallel adder". This work gives a fully specied background dening properly what is a redundant number system and gives tools to help people prove what they have ever thought without having it proved yet. 1 Introduction In this paper, a set of denitions describing general real number representation systems is presented. Our purpose similar to Wiedmer's (see [13]) is to nd a suciently wide denition including classical systems (signed-digit notation, linear numeration systems, p-adic numbers, symmetric level index arithmetic...) but specic enough to make it possible to build some practical results. We focuse on the redundancy property and the relationships between redundancy, on-line and digit-parallel calculus. The most famous redundant system was introduced by Avizienis using a signed-digit representation (see [2]) and led to the denition of on-line arithmetic. On-line operators compute their result digit by digit, producing a new output digit each time they have received a new digit of every operand, most signicant digit rst. The operator computes the rst output digit after having received the rst + 1 digits of every operand: is called the delay of the on-line operator (For more details, the reader can refer to [3, 6]). 2 Getting into the subject Denition 1 A system representing a subset of the real numbers is given by a couple (C F) where C is a nite alphabet and F apartially dened mapping: F : CN ;! R (N is the set of the natural integers and R the set of the real numbers) such that the image of F is dense in a neighborhood of zero inr. The operators +, -, * working on strings of CN are dened withrespect to the following condition: F (X op Y )=F (X) opf (Y ) each time F (X) opf (Y ) belongs to F (CN ). Letters of the alphabet C are denoted byc 0 c 1 ::: y This work has been partially supported by the french "reseau doctoral en architecture des ordinateurs".

4 Denition 2 For any integer n we denote by R n (a 0 a 1 :::a n ) the set of real numbers that can be represented by a string of CN beginning with the n +1 letters a0 a 1 ::: a n. In the following, we focuse on systems where : (H1) For all natural integers n, for all innite string (a i ) i2nr n (a 0 :::a n ) is bounded. (H2) lim n!+1 (Sup(R n(a 0 :::a n )) ; Inf((R n (a 0 :::a n )) = 0. We call them converging systems. These two restrictions are quite natural since the user expects to get more information on the locationofareal number by getting more letters (digits) of one of its representations. It also implies that only a bounded set of reals can be represented. As opposed to converging systems, one can dene diverging systems where the R n (a 0 :::a n ) are notbounded. P adic number systems are diverging systems (see [8,1]). Lemma 1 In a converging real number system, for all natural integer n, and for all innite string (a i ) i2n, R n (a 0 :::a n ) is a closed set in the usual topology of real numbers. In particular: Sup(R n (a 0 :::a n )) = Max(R n (a 0 :::a n )) and will be noted Max(a 0 :::a n ). Proof. We prove that for any sequence (x m ) m2nof R n (a 0 :::a n ) that goes to the real x when m goes to innity, the limit x belongs to R n (a 0 ::: a n ). For each m 2 N let us dene X m 2 CN so that: X m =(a m 0 ::: a m n :::) anda m 0 ::: a m n = a 0 ::: a n. F (X m )=x m. Let us consider the set A = fa m n+1 jm 2 Ng. AsA is included into C, the cardinal of A is nite and there exists c i 2 C so that cardfm 2 Nja m n+1 = c i g =+1. We set then a n+1 = c i and apply this process iteratively by incrementing n and by only taking into account the terms of the sequence that have a 0 ::: a n as n rst letters. This is an usual diagonal process that extracts from the sequence (X m ) m2na new sequence (Y m ) m2n. The following equalities hold: 8m 2 N F(Y m ) 2 R n (a 0 :::a n ) lim m!+1 F (Y m)=x 8m 2 N 8i m Y m i = a i Let us call Y =(a 0 ::: a n :::) andy = F (Y ) the associated real number. We prove in the following that y equals x. We immediatly deduce that x 2 R n (a 0 ::: a n ). For n + 1 letters c 0 ::: c n, let us call diam(c 0 ::: c n ) the real number equal to Sup(R n (c 0 :::c n )) ; Inf(R n (c 0 :::c n )): For any 2 R with >0, let us choose p 2 N so that 8n p diam(a 0 ::: a n ) 1.We obtain : 2 jf (Y ) ; F (Y q )j 1 2 (1) Now let us choose q p so that: jx ; F (Y q )j 1 2 (2) From (1) and (refequ2) we deduce that: jx ; F (Y )j = jx ; yj As this is true for any, x equals to y. 2

5 Denition 3 A number system (F,C) is fully redundant if and only if there exists a non negative integer k that veries for each real number A =(a 0 a 1 :::) and for each integer n k: If Sup(a 0 :::a n ) < Sup(F (CN )) then there exists a string (b0 ::: b n ) dierent from (a 0 :::a n ) (9i with 0 i n and b i 6= a i ) such that Sup(a 0 :::a n ) >Inf(b 0 :::b n ) and Sup(b 0 :::b n ) > Sup(a 0 :::a n ). If Inf(a 0 :::a n ) > Inf(F (CN )) then there exists a string (c0 ::: c n ) dierent from (a 0 :::a n ) such that Sup(c 0 :::c n ) >Inf(a 0 :::a n ) and Inf(a 0 :::a n ) >Inf(c 0 :::c n ). This denition makes the link between redundancy and the overlapping of the representation sets: a real number system is redundant if and only if given a representation (a 0 :::) of a real number, each prex of this representation has got a sux whose combination can be rewritten with a dierent prex. Lemma 2 If in a real number system, one can add numbers with an online delay, then for all integers n>0 there exists (z 0 z 1 :::) a string representing zero so that R n (z 0 z 1 :::z n ) 6= f0g. Proof. Let us suppose that there exists n so that for all string representing zero, R n (z 0 :::z n )=f0g and reduce it to the absurd. Let us take x arealnumber represented by a string X, and;x a string representing its opposite. Let us dene X 0 having the same rst n +1+ digits as ;X and anything after. The rst n + 1 digits of the on-line sum of X and X 0 must be the rst n + 1 digits of a string representing zero. As R n (z 0 :::z n )=f0g we obtain X = X 0. As a result, our number system can at most represent only (card(c)) n++1 real numbers which is absurd because F (CN ) includes a non empty neighborhood of zero. Lemma 3 If in a real number system, one can add numbers with an on-line delay, then for all natural integer n>0 there exists Z =(z 0 z 1 :::) a string representing zero so that R n (z 0 z 1 :::z n ) includes a stricly positive real number and a string Z 0 so that R n (z 0 0 z0 1 :::z0 n) includes a stricly negative real number. Proof. Let us take an integer n>0, from Lemma 2 we produce an innite string representing zero Z =(z 0 z 1 :::) so that there exists Z 0 2 R n+ (z 0 z 1 :::z n+ )nf0g. Let us call Z 00 the opposite of Z 0 and Z 000 the string resulting from the on-line addition of Z 0 and Z 00. Let us consider Z + Z 00,by construction its rst n + 1 digits are the rst digits of a string representing zero (They are the same as those of Z 000 ) and the sign of Z + Z 00 is the opposite of the sign of Z 0. Lemma 4 If in a real number system, one can add numbers with an on-line delay, then for all integer n>0 there exists a string Z =(z 0 z 1 :::) representing zero so that R n (z 0 z 1 :::z n ) includes a strictly positive real number and a strictly negative number. Proof. Let us take an integer n>0, from Lemma 3 we produce two innite strings representing zero Z =(z 0 z 1 :::) andz 0 =(z 0 0 z0 1 :::) so that there exist P 2 R n+(z 0 z 1 :::z n+ ) with P > 0and N 2 R n+ (z 0 0 z0 1 :::z0 n+) with N <0. The rst n + 1 digits resulting from the on-line addition of P + Z 0 and of Z + N are the same and are the rst n + 1 digits of an innite string representing zero. We immediately see that P + Z 0 = P>0 and Z + N = N<0. 3

6 Theorem 1 If in a converging real number system, one can add numbers with an on-line delay then this system is fully redundant. Proof. Let us take a representable real number A =(a 0 a 1 :::) and an integer n > ksuch that Max(a 0 :::a n ) < Max(F (CN )) and Min(a0 :::a n ) > Min(F (CN )). Thanks to Lemma 4, let us choose Z 0 > 0 and Z 00 < 0 with Z 0 and Z 00 2 R n++i (z 0 :::z n+ ). As the system is converging, by suciently increasing the integer i, we canchoose strings Z 0 and Z 00 so that Max(a 0 :::a n )+Z 0 and Min(a 0 :::a n )+Z 00 be representable. Let us call A n = Max(a 0 :::a n ). Let us consider the innite string (b 0 b 1 :::) resulting from the addition of A n and Z 0 : this real number is strictly greater than A n therefore the string (a 0 :::a n ) is dierent from the string (b 0 :::b n ). Moreover, when adding serially A n and Z, we obtain the rst n + 1 digits (b 0 :::b n ) too, so the innite string A n + Z is another representation of A n. The rst n + 1 digits resulting from the on-line addition of A n and Z 00 are (b 0 :::b n ) and A n + Z 00 <A n then: We also verify that: Max(a 0 a 1 :::a n ) >Min(b 0 b 1 :::b n ) Max(b 0 b 1 :::b n ) A n + Z 0 >A n = Max(a 0 a 1 :::a n ) We proceed symmetrically with B n = Min(a 0 :::a n ) and the theorem is proven. Denition 4 An operator that performs the exact addition of two real numbers by deducing the i th digit of the result from the values of at most p positions in the representation of the operands and satisfying the following constraint is called a parallel adder: Let us denote Pos(i A B) the set of the positions in the representation of the operands A and B that have an inuence onthei th digit of the writing of the result of the addition of A and B. Then for all i, the set: fpos(i A B) j A and B are representable.g is bounded by the natural number f(i) (f() 2 NN ). If in a number system there exists an online adder, there exists obviously a parallel adder too. Lemma 5 If in a real number system, a parallel adder can be built then for all natural integer n>0 there exists a string (z 0 z 1 :::) representing zero sothatr n (z 0 z 1 :::z n ) 6= f0g. Proof. Let suppose that there exists n so that for all string representing zero, R n (z 0 :::z n )=f0g and reduce it to the absurd. Let us denote M n = Maxf Pos(i A B) j A and B be representable and 0 i n g. By construction, we have M n Maxf f(i) with0 i n g. Let us take a representation X of a real number and a representation ;X of the opposite. Let us build the innite string X 0 obtained by concatenating the rst M n digits of X and anything after. The rst n + 1 digits of the parallel adding of X 0 and ;X are those of a representation of zero. As R n (z 0 :::z n )=f0g, wegetx 0 = X. At most (card(c)) Mn dierent realnumbers can be represented in our system what is absurd. Lemma 6 If in a real number system, a parallel adder can be built then for all natural integer n>0 there exists Z =(z 0 z 1 :::) a string representing zero sothatr n (z 0 z 1 :::z n ) includes a stricly positive real number and a string Z 0 so that R n (z 0 0 z0 1 :::z0 n ) includes a stricly negative real number. 4

7 Lemma 7 If in a real number system, a parallel adder can be built then for all natural integer n>0 there exists Z =(z 0 z 1 :::) a string representing zero so that R n (z 0 z 1 :::z n ) includes a strictly positive real number and a strictly negative number. Proof. The proofs of Lemmas 6,7 are the same as those of of Lemmas 3,4 by replacing everywhere n + by M n where: M n = Maxf Pos(i A B) j A and B be representable and 0 i n +1g. Theorem 2 If in a converging real number system one can build a parallel adder, then this system is fully redundant. Proof. Let us take a representable real number A = (a 0 a 1 :::) and an integer n > k so that Max(a 0 :::a n ) < Max(F (CN )) and Min(a0 :::a n ) > Min(F (CN )). Thanks to Lemma 7, let us choose strings Z 0 > 0andZ 00 < 0withZ 0 and Z 00 2 R Mn+i (z 0 :::z Mn+i ). As the system is converging, by suciently increasing the integer i, we canchoose Z 0 and Z 00 so that Max(a 0 :::a n )+Z 0 and Min(a 0 :::a n )+Z 00 be representable. The end of the proof is identical to the end of the proof of Theorem 1. Theorem 3 If in a converging real number system, one can multiply numbers with an on-line delay in a non-empty interval including 1, then this system is fully redundant. Proof. We proceed as for Theorem 1 by proving these successive lemma: Lemma 8 If in a number system, one can multiply numbers around one with an online delay, then for all integers n>0 there exists a string representing one (u 0 u 1 :::) such that R n (u 0 u 1 :::u n ) 6= f1g. Lemma 9 If in a real number system, one can multiply numbers around one with an on-line delay, then for all integers n>0 there exists a string representing one U =(u 0 u 1 :::) such that R n (u 0 u 1 :::u n ) includes a real number strictly greater than one and a string U 0 so that R n (u 0 0 u0 1 :::u0 n) includes a real number strictly less than one. Lemma 10 If in a real number system, one can multiply numbers around one with an on-line delay, then for all integers n>0 there exists a string representing one U =(u 0 u 1 :::) such that R n (u 0 u 1 :::u n ) includes a real number strictly greater than one and another strictly less than one. One can easily deduce the proofs of these three lemmas from the proofs of Lemmas 2,3,4 by replacing everywhere the real number zero by one, the on-line addition by the on-line multiplication and the opposite by the inverse. By noticing that adding a small positive number to a real number can be viewed as multiplying it by anumber a little greater than one, the proof of Theorem 1 stands for these theorems too. Denition 5 In a fully redundant and converging number system, let us call Rd n and rd n (the maximal and minimal power representation at rank n) the values: Rd n = Maxf Max(a 0 :::a n ) ; Min(a 0 :::a n ) for all string (a 0 :::a n ) g rd n = Minf Max(a 0 :::a n ) ; Min(a 0 :::a n ) for all string (a 0 :::a n ) g 5

8 Denition 6 A syntactically dense real number system satises: for each A a representable real number and for each integer n, the sets R n (a 0 :::a n ) are closed intervals. In such a system, we call M(a 0 :::a n ) the middle of the interval R n (a 0 :::a n ) and I = F (CN ) the interval of the representable real numbers. Denition 7 In a fully redundant, syntactically dense system, let us take (a 0 :::a n ) arepresentation of any representation interval I A (We will use the abusive notation I A = R n (a 0 :::a n ) in the following) then the writing I B = R n (a 0 :::a n;1 b n ) is a right neighbor of I A if and only if: Max(a 0 :::a n;1 b n ) >Max(a 0 :::a n ). For all writing I C =(a 0 :::a n;1 c n ) so that Max(a 0 :::a n;1 c n ) >Max(a 0 :::a n ),wehave: Min(a 0 :::a n;1 b n ) Min(a 0 :::a n;1 c n ). The same denition holds for left neighbor by replacing maxima by minima and >'s by <'s and conversely. By extension, a representation interval I B is a right (respectively left) neighbor of arepresentation interval I A if and only if I B has got a writing that is a right (respectively left) neighbor of a writing of I A. By construction, the sets R n (a 0 :::a n ) are covering I. A representation interval of rankn(withn +1 digits) is principal if and only if it is not included in one of its neighbors. Two representation intervals are said to be neighbors if the rst one is a left neighbor of the second one and the second one a right neighbor of the rst one. In the following, we will call redundancy degree atrank n the value d n equal to the length of the smallest intersection between two neighbors. Lemma 11 Let us consider two representation intervals I A = R n (a 0 :::a n ) and I B = R n (a 0 :::a n;1 b n ) so that I B is a right neighbor of I A and I A is principal, then I A is a left neighbor of I B. Proof. By construction, Min(a 0 :::a n ) <Min(a 0 :::a n;1 b n ) (if not, I A is included in I B what denies the fact that I A is principal). Let us take I C =(a 0 :::a n;1 c n ) so that Min(a 0 :::a n;1 c n ) <Min(a 0 :::a n;1 b n ). Let us suppose that Max(a 0 :::a n;1 c n ) >Max(a 0 :::a n ). This is equivalent to assuming that I A is not a left neighbor of I B and let us reduce that hypothesis to the absurd: First case: Min(a 0 :::a n;1 c n ) >Min(a 0 :::a n ) and in this case, I C is a right neighbor of I A what refutes the fact that I B is one: absurd. Second case: Min(a 0 :::a n;1 c n ) Min(a 0 :::a n ) then I A is strictly included in I C what is absurd since I A is principal. To conclude, I A is a left neighbor of I B. Lemma 12 Let us consider a fully redundant and syntactically dense system where for all natural integer n, the redundancy degree atrank n satises d n > 0. Let us take a principal representation interval I A =(a 0 :::a n ) so that Max(a 0 :::a n ) <Max(a 0 :::a n;1 ) (respectively Min(a 0 :::a n ) > Min(a 0 :::a n;1 )), then there exists I B = (a 0 :::a n;1 b n ) so that the length of the intersection of I A and I B is at least d n and Max(a 0 :::a n ) < Max(a 0 :::a n;1 b n ) (respectively Min(a 0 :::a n ) > Min(a 0 :::a n;1 b n )). 6

9 Proof. I A =(a 0 :::a n ) is principal and Max(a 0 :::a n ) <Max(a 0 :::a n;1 ). Let us denote E the set of the representation intervals R n (a 0 :::a n;1 c i ) so that Max(a 0 :::a n;1 c i ) >Max(a 0 :::a n ). This set is not empty otherwise Max(a 0 :::a n;1 )would be equal to Max(a 0 :::a n ). Let us choose one of the elements of E with the smallest lower bound: I B = R n (a 0 :::a n;1 c j ). The intersection of I A and I B is not empty otherwise the set of the intervals R n (a 0 :::a n;1 c i ) does not cover R n (a 0 :::a n;1 ). So by denition, I B is a right neighbor of I A, thanks to Lemma 11 I A is a left neighbor of I B then by denition of the redundancy degree, their intersection is at least of length d n.we proceed in a similar fashion for the minima. It is useful to notice that this result holds when n equals zero too if we decide that R n (a 0 :::a ;1 )= F (CN )=I. Theorem 4 In a fully redundant and syntactically dense system, if for all natural integer n 0 0 <Rd n+ 1 2 d n then it is possible to perform on-line addition with delay. By applying this theorem to the well-known signed-digit representation systems (see [2]) we obtain an online adder of delay 2 in radix 2 and an online adder of delay 1 in higher radices: these are known to be the best delays. The proof builds an eective online algorithm of addition that can be viewed as the most general online addition algorithm possible. It gives upper bounds for the delay of the online adder in \exotic" number systems (For example radix with digits in the set f;2 ; g). In the following, a limited converse proposition will be presented. Proof. We show by induction the following hypothesis P (n): For all n 0, for all representable real numbers A and B whose sum is also representable, we suppose that the rst n + 1 digits of a representation of the sum c 0 ::: c n have already been produced, when the digits a n++1 and b n++1 are available, it is possible to nd a digit c n+1 so that A + B 2 R n+1 (c 0 :::c n+1 ) First let us prove the hypothesis P (0). Let us dene the real number C = M(a 0 :::a )+M(b 0 :::b ). Let us call E the set of letters (digits) c i of the alphabet so that C 2 R 0 (c i ). First case: E is empty. This is possible only if C does not belong to I (since the system is syntactically dense). Let us suppose that C>Max(I) (The other case is similar). Let us take c j so that Max(c j )=Max(I) and R 0 (c j ) is principal. If R 0 (c j )=I then A + B 2 R 0 (c j ) and the new digit to produce is c j. In the other case, thanks to Lemma 12, there exists a digit c k so that the length of the interval R 0 (c j ) \ R 0 (c k ) be greater or equal to d 0. Let us consider the scheme: A+B M(c j ) C As jc ; (A + B)j 1 2 (Max(a 0:::a ) ; Min(a 0 :::a )+Max(b 0 :::b ) ; Min(b 0 :::b )) jc ; A + Bj Rd p then we deduce : jc ; (A + B)j 1 2 d 0 (3) 7

10 A fortiori jmax(c j ) ; (A + B)j < 1 2 d 0 and A + B 2 R 0 (c j ) The algorithm produces the digit c j. QED Second case: Let us take R 0 (c j ) a principal interval so that C 2 R 0 (c j ). Let us suppose that Max(c j ) ; C C ; Min(c j ) (In the other case, the proof is identical if we replace each right neighbor by a left neighbor.) then: C ; Min(c j ) 1 2 d 0 (4) Let us call this step in the proof the step Look for right neighbor. { If R 0 (c j ) does not have any right neighbor, we immediatly deduce from Lemma 12 that Max(c j )=Max(I) and with (3) it is obvious that A + B 2 R 0 (c j ) QED. A+B j M(c ) C Max(I) { Let us call R 0 (c k ) a principal interval which is a right neighbor of R 0 (c j ). If C does not belong to R 0 (c k ) the digit c j is produced: A+B j M(c ) C M(c ) As the length of the intersection of the intervals R 0 (c j ) and R 0 (c k ) is greater or equal to d 0, With (4), we deduce that A + B 2 R 0 (c j ) QED. If C belongs to R 0 (c k ) and Max(c j ) ; C C ; Min(c k ), the digit c j is produced. k A+B M(c j ) C k M(c ) We immediatly deduce that Max(c j ) ; C 1 2 d 0 and with (4) A + B 2 R 0 (c j ) QED. If C belongs to R 0 (c k ) and C ; Min(c k ) >Max(c j ) ; C: Then we consider the interval R 0 (c k ) instead of R 0 (c j ) and restart the process at the step Look for right neighbor. We easily check that equation (4) is always true after the substitution. 8

11 A+B? j M(c ) C M(c ) Since the process never enters an innite loop (The right neighbor substitution can be called at most cardinal(alphabet) times), we are able to produce in any case a digit c j so that A + B 2 R 0 (c j ) with the rst + 1 digits of A and B and the recurrent hypothesis P (n) is proved at step n =0. k Now, let us suppose that the hypothesis P (n) is true for an n 0. Let us prove P (n + 1): with the rst n digits of A and B, the rst n + 1 digits of a representation of A + B are already produced: d 0 :::d n. Let us consider C = M(a 0 :::a n++1 )+M(b 0 :::b n++1 ). Let us call E the set of letters (digits) c i of the alphabet such that C 2 R 0 (c i ). We develop exactly the same cases as for n = 0. The proofs are strictly identical except for the interval I which is replaced everywhere by R n (d 0 :::d n ). By recurrence, the hypothesis is proved for all natural integer n. Denition 8 In a redundant real number system an adder is said to be reversible if and only if for all A =(a 0 :::) and B =(b 0 :::) two representations of the same real number, there exists an innite string Z A B representing zero sothata + Z A B = B. Theorem 5 In a fully redundant syntactically dense number system, if there exists n so that rd n+ >d n, then no algorithm can perform a reversible addition in delay. Proof. Let us take n so that rd n+ >d n. Let us choose I A =(a 0 :::a n )andi B =(a 0 :::a n;a b n )so that: I A be a left neighbor of I B. I B be a right neighbor of I A. The length of the intersection of I A and I B is equal to d n. Let us denote A = Max(I A )=(a 0 :::a n :::a n+ :::) =(a 0 :::a n;1 b n :::) andz A B =(z 0 z 1 :::). Let us suppose that it is possible to perform addition in delay and reduce it to the absurd. a 0... a n+d... z 0... z n+d... a 0... a n-1b n... If we add serially Min(a 0 :::a n+ ) and Min(z 0 :::z n+ ), we obtain a real number coded by a string beginning with (a 0 :::a n;1 b n )too. We immediatly deduce that the length of I A \ I B is greater or equal to rd n+ which refutes the fact that rd n+ >d n. 9

12 Denition 9 By slightly modifying the previous set of denitions, one can adapt them to describe relative and natural integer number systems: one works only on nite length strings instead. A system would be converging if the sets R n (a 0 :::a n ) are included in a strictly length decreasing intervals sequence (I n ) n2n: for 0 n n max R n (a o :::a n ) I n and R n (a o :::a n ) 6 I n+1 I 0 I 1 ::: I nmax and I nmax contains only one integer. The redundancy denition does not change and we obtain the same results concerning redundancy and on-line addition with delay in relative integer number systems. A relative or natural integer number system is syntactically dense if the R n (a 0 :::a n ) are intervals of Z or N. Theorems 4 and 5 still holds for relative and natural integers: one only has to be cautious to nd the correct last digits since the algorithm does not produce them but still prove that they exist. Lemma 13 This is to show how to use Theorem 4 in a practical case. For M any large natural integer, we dene the encoding function F from f0 1g M+1 into N by: F (a 0 :::a M )= P M i=0 a iu M;i where U i is the Fibonacci sequence with U 0 =1and U 1 =2. This system is called the Fibonacci Numeration system (see [5]). In this system, it is possible to perform an on-line addition with delay 3. Proof. The fact that this system P is syntactically dense derives from the normalization algorithm M that nds for eachinteger in f0 P U i=0 ig a canonical P representation P by using the Euclidean division. n By denition, R n (a 0 :::a n )=[ a n i=0 i U M;i a M;n;1 i=0 i U M;i + i=0 U i ]. From that we immediatly deduce that this system is fully redundant and that: Rd n (a 0 :::a n )=rd n (a 0 :::a n )= X X M;n;1 U i i=0 M;n;1 d n =( U i ) ; U N;n;1 = Rd n+1 i=0 Thanks to Theorem 4, it is possible to produce the rst M ; 2 digits of a representation of the sum with an on-line adder of delay 3 (when the M + 1 digits of both operands have been serially supplied) if for all n 2f0 ::: M; 3g: That is to say 8n 2f0 ::: M ; 3g, That is equivalent to8n 2f0 ::: M ; 3g: Rd n d n (5) Rd n Rd n+1 Rd n+1 2U M;n;1 By replacing n +1by M ; i, 8i 2f2 ::: M ; 1g: We prove that proposition by induction on i: Rd M;i 2U i First Rd M;2 = U 0 + U 1 2U 2 10

13 Let suppose the proposition true at step i: Rd M;(i+1) = Rd M;i + U i 2U i + U i;1 + U i;2 2U i +2U i;1 =2U i+1 QED. At the step M, the on-line adder has got all the digits of the operands then it can produce the three last digits of the sum. As a matter of fact, one can easily adapt the general on-line adding algorithm to nd a nite automaton that adds in delay 3 in the Fibonacci Numeration system. 3 Conclusion In this paper, we have presented a set of denitions from which one can derive a classication of non-classical number representation systems (see gure 1). Details on the number systems can be found in the corresponding articles [2, 3,4,5,7,8,10,11,12]. Finite Alphabet Infinite Alphabet Diverging Diverging Converging Least significant digit first systems P_adique numbers Converging Most significant digit first systems Mixed Radix systems Linear numeration systems Sign/logarithmic number system Residue Number Systems Continued fractions SLI arithmetic Figure 1: Real number and numeration system classication. From a general point of view, converging systems are characterized by the fact that the usual arithmetic operations are expensive (At least, addition or multiplication is logarithmic compared to the size of the operands) but they allow serial computations and comparison algorithms are straightforward. On the other hand, in diverging systems the arithmetic operations can be very ecient (even constant time in residue arithmetic) but comparisons are very costly and serial calculus can not be achieved for division. The dream of every computer arithmetician is to break the wall between these two classes and to build an hybrid arithmetic that only keeps the advantages of both sides. The exploration path we have chosen is the one of redundancy. The reasons are multiple: by accepting to lose some information on numbers we gain a freedom degree that allows to build more ecient algorithms at the long run. Thanks to redundancy, it is possible to build systems where numbers ow serially digit by digit which is the only way to handle large numbers without a tremendous hardware cost. Another point is how far we are able to represent \real" real numbers. Most systems only represent and handle a bounded set of rational numbers although the recursive real numbers set a theoretical far-beyond frontier (see [9]). Redundancy by allowing serial calculus on converging systems is the key to \practical exact" real arithmetic. (No hope to obtain a practical exact real arithmetic on a diverging system since one may need an arbitrary large number of digits to only know if the result is close to zero.) We will conclude by giving some 11

14 conjectures whose conrmation or invalidation would no doubt lead to the design of useful tools and results for computer arithmetic: Let us consider an exact real number system with a nite alphabet that can represent real numbers of any magnitude, then no nite automaton can perform serially addition on such a system. In no integer number system, one can as well perform addition, multiplication and comparison with nite automata. 12

15 References [1] Y. Amice Les nombres p-adiques. Presses Universitaires de France, Col. Sup., [2] A. Avizienis Signed-digit number representations for fast parallel arithmetic. IRE Trans. Electron. Comput. 10, pp 389{400, Sept [3] M.D.Ercegovac and K.S.Trivedi On-line algorithms for division and multiplications IEEE Transactions on Computers,Vol. C-26 N. 7, pp , July [4] A. S. Fraenkel Systems of numeration The American Mathematician Monthly, pp 105{114, Feb [5] C. Frougny Representation of numbers in Non-Classical Numeration Systems. Proc 10th IEEE Symp. on Comp. Arith., pp 17{21, June [6] A.Guyot,Y.Herreros and J.M.Muller JANUS, An on-line multiplier/divider for manipulating large numbers. 9 th Symposium On Computer Arithmetic, Santa-Monica, Sept [7] D. E. Knuth The Art of Computer Programming, vol 2, Seminumerical algorithms, pp 178{364, Addison Wesley, [8] E. V. Krishnamurthy, T. M. Rao, K. Subramanian Finite-segment p-adic number systems with applications to exact computations. Proc. Indian Acad. Sci., vol. LXXXI, pp 58{79, [9] H. G. Rice Recursive real numbers. Proc. Americ. Math. Soc., vol. 5, no. 5, pp 784{791, [10] E. E. Swartzlander and A. G. Alexopoulos The sign/logarithmic number system IEEE Transactions on Computers, vol C-24, pp 1238{1242, Dec [11] P. E. Turner Will the "Real" Real Arithmetic Please Stand Up? Comp. and Math., vol 38, Number 4, pp 298{304, April [12] J. E. Vuillemin Exact Real Computer Arithmetic with Continued Fractions IEEE Transactions on Computers, vol 39, Number 8, pp 1087{1105, Aug [13] E. Wiedmer Computing with innite objects Theor. Comput. Sci., vol 10, pp. 133{155,